AI News

HyperWorld studies how state serialization affects learned textual world models for language-model agents. Entity-centered hyperedge units provide the clearest gains for 0.5B–1.5B models and under distribution shift, while achieving the highest downstream greedy-planning success rate. Larger models reduce the difference, and pairwise triples can remain competitive on in-distribution exact match.

IMPACT addresses a supervision imbalance in world model training by using cross-attention and local prediction errors to focus denoising on manipulated-object regions. Tests across robot-arm and human-hand manipulation show better interaction fidelity, physical plausibility, and visual quality than MSE baselines.

Researchers introduce WM-R1, a reinforcement-learning framework that trains mobile GUI agents with world models instead of real Android environments. World models generate state transitions during rollouts and help agents reason about candidate action consequences before execution. The approach removes real-environment interaction, supports massively parallel and step-level trajectory generation, and uses a multi-dimensional rule-based reward for task success, efficiency, and world-model utilization. On Android benchmarks, agents trained with WM-R1 significantly outperform GRPO-only baselines and inference-time simulation methods. The training dataset contains 2,000 challenging tasks, and the implementation is publicly available.